Solutions/AI
Artificial intelligence,
running on your own infrastructure.
We build enterprise AI platforms on Red Hat OpenShift AI: from data preparation to model serving, with governance and data that stays inside your perimeter. This area is still being defined — the content below is demonstrative.
Overview
From pilot to production,
without moving your data out.
Most AI initiatives stall at the pilot stage: models never reach production, data is scattered, and compliance requirements block public AI services. Our approach starts with the platform, not the model.
Services
Enterprise AI platform
Red Hat OpenShift AI with GPU scheduling, tenant isolation and federated access.
MLOps & model serving
Automated pipelines for training, validation and delivery of models into production.
Data engineering
Ingestion, cleaning and data governance, with catalogue and access policies.
AI in operations
Anomaly detection across telemetry and logs for critical systems.
AI in security
Event correlation and automatic alert prioritisation in the SOC.
Enablement
Training, reference architecture and support for internal teams.
Architecture
we use
Full stack → Metaminds runs specialised AI agents and assistants on the same governed platform as everything else: local open-weight models on our own GPUs for sensitive data, Claude via API where data policy permits, tools exposed only through MCP, and every action landing as a reviewable Git change.
↓ Requests flow down — context and tools attached ↑ Answers & actions flow up — via Git, traced end-to-end
Ask, draft, review and approve from Slack, Backstage or the IDE
Tenders, proposals, as-built documentation
Public sector and regulated industries, isolated per tenant
AIOps agents reading telemetry, acting through GitOps
Bid & tender analysis: requirements, risks, go/no-go draft
Turns live configuration into as-is documentation
Proposals, as-built docs, code and review support
Read-only Q&A grounded in the platform catalog
Correlate alerts, gather context, draft fixes as MRs
Drives promotions and rollouts, rolls back on bad signals
Tool-using agents with permissions and session state
Multi-step flows, retries, human checkpoints
The only way agents touch systems: typed, scoped, audited
Docs, configs, tickets and runbooks as context
Conversations, task state, approvals, outcomes
Least-privilege identities, admission policy, secrets by reference
Routes by data class; quotas, redaction, logging
Chat, reasoning and Romanian document intelligence on our GPUs
Top-tier reasoning and coding where policy allows
Feed the retrieval layer; nothing leaves the site
Serving, notebooks, registry, fine-tuning; signed model artefacts
Traces, cost, latency, quality; evals gate autonomy
Accelerated compute on OpenShift GPU node pools, scheduled and partitioned
Multi-site, multi-tenant, GitOps-driven; agents deploy like any workload
What agents read — and are measured by
Sovereign; data residency by design
Governance
Sovereign by design
Increasing autonomy never means decreasing control — the guardrails are the platform’s own.
Design principles
How the AI layer behaves
Augmentation, not abdication: engineers move up to supervision, judgement and design.
Why it matters
Outcomes
Sovereign AI, credibly
Inference, retrieval and memory on our own infrastructure — a real answer for public sector and regulated tenants.
Faster bids, docs and delivery
Claudia, Dobby and Uzi already remove hours from tenders, documentation and proposals.
Auditable autonomy
Every agent action is a traced call and a Git change — explainable to auditors and customers.
One platform, not two
AI is a workload on the existing platform, inheriting its HA, security and GitOps discipline.
